图像不变局部特征研究及应用
文献类型：学位论文
| 作者 | 黎俊 |
| 学位类别 | 博士 |
| 答辩日期 | 2009-01-14 |
| 授予单位 | 中国科学院软件研究所 |
| 授予地点 | 软件研究所 |
| 关键词 | 不变局部特征 摄像机运动检测 目标模型 目标识别 |
| 其他题名 | Research and Application of Local Invariant Image Features |
| 中文摘要 | 图像不变局部特征是新近兴起的一类图像特征，基于不变局部特征的图像表示是计算机视觉的热点研究问题，在理论研究和实际应用上都具有重要意义。本论文针对图像不变局部特征的原理特性及应用展开研究：（1）当今流行的不变局部特征检测和描述方法；（2）局部特征组织方式；（3）基于局部不变特征的摄像机运动检测方法；（4）基于局部特征组合的目标模型及识别方法。 深入研究了当今流行的不变局部特征检测子，重点分析了其提取原理、特征结构、不变性阶次、精确度等特性，在此基础上对多种检测子进行比较分析，得出各自的适用范围，并总结出在具体应用环境下的特征选择原则。 针对视频分析中摄像机运动检测的具体应用，提出一种基于尺度不变局部特征的摄像机运动检测方法。该方法选取尺度不变局部特征，采用无序特征集合的方式表示图像帧，通过帧间局部特征的匹配，提出归一化软投票的方法鲁棒地估计特征匹配对的位置、尺度的变化，并根据各变化值和投票数的特点识别出摄像机的运动类型。该方法简单、鲁棒，满足了摄像机运动检测的处理速度和准确性需求。 针对基于局部特征的目标表示和识别问题，研究分析现有两种模型bag-of-words和part-based的优缺点，将二者结合，提出一种局部特征组合的目标表示模型和相应的识别算法。该方法在半局部区域内的特征同时进行外观描述和空间位置编码，并用数据挖掘中的频繁项挖掘技术自动提取出表征目标的特征组合，作为子模型。目标模型由一系列子模型构成，子模型的数量及每个子模型中包含的部件数目均自动从训练集中发现，是完全目标自适应的。所提方法克服了bag-of-words方法表达的精确性不足、part-based方法训练速度过慢的缺点，在识别问题上得到了较好的总体性能。 |
| 英文摘要 | As newly developed image features, local invariant features are becoming more and more popular in image representation and object recognition tasks in the field of computer vision and pattern recognition. This dissertation mainly focuses on the following aspects: (1) Research of the most popular invariant feature detectors and descriptors; (2) Organization of local features; (3) Camera motion detection based on scale-invariant local features; (4) Object recognition based on combinations of local invariant features. The most important characteristics of local invariant feature detectors, including feature structure, level of invariance, preciseness of feature position, are deeply investigated. And several popular detectors are compared accordingly. Principles on how to choose appropriate features for specific application are summarized. A new camera motion detection method based on scale invariant local features is proposed. Frames are presented as a bag of scale-invariant local features. And a soft matching scheme is designed to estimate the change of position and scale between matched pairs. The estimated changes in position and scale and corresponding votes can be used as features and identify camera motion type robustly. The proposed method is proved to be both efficient and robust. In the task of object recognition in real scenes, the pros and cons of two current classic models, bag-of-words model and part-based model, are analyzed; and a new object model, which combines the two classic models, is proposed. The model is composed of several sub-models, and each sub-model is a frequent combination of local features in a neighborhood. The number of sub-models and the number of features in each sub-model are automatically discovered by the efficient frequent item mining (FIM) method from training set, making the object model fully adaptive. Experiments indicates that the new model achieved more precise description of the object than the bag-of-words model and much faster training speed than the part-based model. |
| 语种 | 中文 |
| 公开日期 | 2011-03-17 |
| 页码 | 65 |
| 源URL | [http://124.16.136.157/handle/311060/7400]  |
| 专题 | 软件研究所_综合信息系统技术国家级重点实验室 _学位论文 |
| 推荐引用方式 GB/T 7714 | 黎俊. 图像不变局部特征研究及应用[D]. 软件研究所. 中国科学院软件研究所. 2009. |
入库方式： OAI收割
来源：软件研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。